Video cameras and digital still cameras generally employ a single image sensor with a color filter array to record a scene. This approach begins with a sparsely populated single-channel image in which the color information is encoded by the color filter array pattern. Subsequent interpolation of the neighboring pixel values permits the reconstruction of a complete three-channel, full-color image. This full-color image, in turn, can be noise-cleaned, sharpened, or color corrected to improve, or enhance, the appearance of the image. This image enhancement can be greatly facilitated by computing an edge map of the image in order to classify the image into edge regions and flat regions. This permits the use of algorithms that perform different computations for edge regions and for flat regions. One popular approach is to either directly detect or synthesize a luminance color channel, e.g. “green”, and then to generate an edge map from the luminance image. U.S. Pat. No. 6,614,474 (Malkin et al.) describes computing a luminance channel and then generating edge information from a set of directional edge detection kernels. The problem with this approach is that edges that vary only in chrominance and not luminance run the risk of being undetected. To address this concern, U.S. Pat. No. 5,420,971 (Westerink et al.) teaches computing a YUV luminance-chrominance image, computing edge information from all three channels (Y, U, and V), and then combining them as an L2-norm to detect both luminance and chrominance edges. The problem with this approach is the noisiness of the computed luminance-chrominance image is defined by the noisiness of the original color data, e.g., RGB. This level of noise in the original color data is determined, among other things, by the relative narrowness of the spectral frequency response of the individual color channels. When the scene being captured is well lit, e.g., a sunny landscape, the narrowness of the spectral frequency responses is usually not an issue. When the scene is not well lit, e.g., indoors, or the exposure time is necessarily short to reduce motion blur, e.g., at a sporting event, the relative narrowness of the spectral frequency response of the individual color channels can produce noisy images.
Under low-light imaging situations, it is advantageous to have one or more of the pixels in the color filter array unfiltered, i.e. white or panchromatic in spectral sensitivity. These panchromatic pixels have the highest light sensitivity capability of the capture system. Employing panchromatic pixels represents a tradeoff in the capture system between light sensitivity and color spatial resolution. To this end, many four-color color filter array systems have been described. U.S. Pat. No. 6,529,239 (Dyck et al.) teaches a green-cyan-yellow-white pattern that is arranged as a 2×2 block that is tessellated over the surface of the sensor. U.S. Pat. No. 6,757,012 (Hubina et al.) discloses both a red-green-blue-white pattern and a yellow-cyan-magenta-white pattern. In both cases, the colors are arranged in a 2×2 block that is tessellated over the surface of the imager. The difficulty with such systems is that only one-quarter of the pixels in the color filter array have highest light sensitivity, thus limiting the overall low-light performance of the capture device.
To address the need of having more pixels with highest light sensitivity in the color filter array, U.S. Patent Application Publication No. 2003/0210332 (Frame) describes a pixel array with most of the pixels being unfiltered. Relatively few pixels are devoted to capturing color information from the scene producing a system with low color spatial resolution capability. Additionally, Frame teaches using simple linear interpolation techniques that are not responsive to or protective of high frequency color spatial details in the image.